> ## Documentation Index
> Fetch the complete documentation index at: https://docs.eigenpal.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Design principles

> The engineering choices behind EigenPal workflows, evaluations, and artifacts.

EigenPal is built around a few engineering constraints that make AI workflows
easier to test, review, and operate.

## 1. Evaluations define correctness

A workflow defines behavior. An [evaluation](/concepts/evals) defines the output
contract for a dataset. Models, prompts, and steps can change; the evaluation is
the check that says whether the output still matches what the workflow is meant
to produce. Weighted evaluators roll up to one score against a pass threshold.

## 2. Production issues become tests

When a real document produces a wrong answer, add that input and corrected
output to the evaluation dataset. From then on, the case is part of every
experiment and comparison. See
[Eval-first development](/guides/eval-first-development).

## 3. Keep humans in the loop with Reviews

[Reviews](/concepts/reviews) let people judge production runs, track open
issues, correct expected output, and monitor quality over time. Human feedback
should not live in Slack threads or spreadsheets; it should become part of the
run history and the improvement loop.

## 4. Optimize against evaluation scores

The goal is to meet the quality bar at the lowest acceptable cost. Once
evaluations define that bar, you can compare OCR engines, prompts, and language
models and keep the cheapest version that still passes. See
[Optimize cost](/guides/optimize-cost).

## 5. Treat AI systems like software

Workflows, datasets, evaluators, and agent source are
[artifacts](/concepts/artifacts): plain files you version in Git, review in pull
requests, export between environments, and generate programmatically. The same
discipline you apply to code applies here.

## 6. Separate definition from execution

What a workflow does is configuration; how it runs is the engine's job. The
definition is portable and transparent, while the stateful, queue-based engine
handles retries, load, and long-running work. Definitions and their history are
exportable.